System & ML Performance

Having investigated the new A13’s CPU performance, it’s time to look at how it performs in some system-level tests. Unfortunately there’s still a frustrating lack of proper system tests for iOS, particularly when it comes to tests like PCMark that would more accurately represent application use-cases. In lieu of that, we have to fall back to browser-based benchmarks. Browser performance is still an important aspect of device performance, as it remains one of the main workloads that put large amounts of stress on the CPU while exhibiting performance characteristics such as performance latency (essentially, responsiveness).

As always, the following benchmarks aren’t just a representation of the hardware capabilities, but also the software optimizations of a phone. iOS13 has again increased browser-based benchmarks performance by roughly 10% in our testing. We’ve gone ahead and updated the performance figures of previous generation iPhones with new scores on iOS13 to have proper Apple-to-Apple comparisons for the new iPhone 11’s.

Speedometer 2.0 - OS WebView

In Speedometer 2.0 we see the new A13 based phones exhibit a 19-20% performance increase compared to the previous generation iPhone XS and the A12. The increase is in-line with Apple’s performance claims. The increase this year is a bit smaller than what we saw last year with the A12, as it seems the main boost to the scores last year was the upgrade to a 128KB L1I cache.

JetStream 2 - OS Webview

JetStream 2 is a newer browser benchmark that was released earlier this year. The test is longer and possibly more complex than Speedometer 2.0 – although we still have to do proper profiling of the workload. The A13’s increases here are about 13%. Apple’s chipsets, CPUs, and custom Javascript engine continue to dominate the mobile benchmarks, posting double the performance we see from the next-best competition.

WebXPRT 3 - OS WebView

Finally WebXPRT represents more of a “scaling” workload that isn’t as steady-state as the previous benchmarks. Still, even here the new iPhones showcase a 18-19% performance increase.

Last year Apple made big changes to the kernel scheduler in iOS12, and vastly shortened the ramp-up time of the CPU DVFS algorithm, decreasing the time the system takes to transition from lower idle frequencies and small cores idle to full performance of the large cores. This resulted in significantly improved device responsiveness across a wide range of past iPhone generations.

Compared to the A12, the A13 doesn’t change all that much in terms of the time it takes to reach the maximum clock-speed of the large Lightning cores, with the CPU core reaching its peak in a little over 100ms.

What does change a lot is the time the workload resides on the smaller Thunder efficiency cores. On the A13 the small cores are ramping up significantly faster than on the A12. There’s also a major change in the scheduler behavior and when the workload migrates from the small cores to the large cores. On the A13 this now happens after around 30ms, while on the A12 this would take up to 54ms. Due to the small cores no longer being able to request higher memory controller performance states on their own, it likely makes sense to migrate to the large cores sooner now in the case of a more demanding workload.

The A13’s Lightning cores are start off at a base frequency of around 910MHz, which is a bit lower than the A12 and its base frequency of 1180MHz. What this means is that Apple has extended the dynamic range of the large cores in the A13 both towards higher performance as well as towards the lower, more efficient frequencies.

Machine Learning Inference Performance

Apple has also claimed to have increased the performance of their neural processor IP block in the A13. To use this unit, you have to make use of the CoreML framework. Unfortunately we don’t have a custom tool for testing this as of yet, so we have to fall back to one of the rare external applications out there which does provide a benchmark for this, and that’s Master Lu’s AIMark.

Like the web-browser workloads, iOS13 has brought performance improvements for past devices, so we’ve rerun the iPhone X and XS scores for proper comparisons to the new iPhone 11.

鲁大师 / Master Lu - AIMark 3 - InceptionV3 鲁大师 / Master Lu - AIMark 3 - ResNet34 鲁大师 / Master Lu - AIMark 3 - MobileNet-SSD 鲁大师 / Master Lu - AIMark 3 - DeepLabV3

The improvements for the iPhone 11 and the new A13 vary depending on the model and workload. For the classical models such as InceptionV3 and ResNet34, we’re seeing 23-29% improvements in the inference rate. MobileNet-SSD sees are more limited 17% increase, while DeepLabV3 sees a major increase of 48%.

Generally, the issue of running machine learning benchmarks is that it’s running through an abstraction layer, in this case which is CoreML. We don’t have guarantees on how much of the model is actually being run on the NPU versus the CPU and GPU, as things can differ a lot depending on the ML drivers of the device.

Nevertheless, the A13 and iPhone 11 here are very competitive and provide good iterative performance boosts for this generation.

Performance Conclusion

Overall, performance on the iPhone 11s is excellent, as we've come to expect time and time again from Apple. With that said, however, I can’t really say that I notice too much of a difference to the iPhone XS in daily usage. So while the A13 delivers class leading performance, it's probably not going to be very compelling for users coming from last year's A12 devices; the bigger impact will be felt coming from older devices. Otherwise, with this much horsepower I feel like the user experience would benefit significantly more from an option to accelerate application and system animations, or rather even just turn them off completely, in order to really feel the proper snappiness of the hardware.

SPEC2006 Perf: Desktop Levels, New Mobile Power Heights GPU Performance & Power
POST A COMMENT

239 Comments

View All Comments

  • Jon Tseng - Wednesday, October 16, 2019 - link

    Nice! Any additional thoughts on the U1 UWB chip. I guess not much you can do with it yet but to me the possibilities are intriguing... Reply
  • Andrei Frumusanu - Wednesday, October 16, 2019 - link

    I think Apple has more plans with it in the future, but yes right now it doesn't do very much. Reply
  • tipoo - Wednesday, October 16, 2019 - link

    Definitely think it's getting the hardware ready for the AR glasses. Hyper precise location tracking just by putting your phone down on a desk and having the U1 chips communicate. Reply
  • Diogene7 - Wednesday, October 16, 2019 - link

    I am dreaming of that the Apple U1 UWB chip could be used in a not too distant future (2020 / 2021) for precise spatial locasization for at (short) distance wireless charging : by knowing where exactly in space an Apple device is, Apple might be able to dynamically and efficiently focus wireless energy transfer maybe through wireless resonant charging (Airfuel) for an iPhone or through RF charging like Energous / Ossia for recharging Apple Airpods from an iPhone...

    I think I am dreaming, but just hope that Apple is working hard to make wireless power at a short distance a reality : I would dream to be able to drop my iPhone anywhere on my bedside table, and that it automatically recharge during the night from a base station up to a distance of 1,5 foot / 50cm : it would bring sooooo much more convenience than Qi wireless charging...
    Reply
  • patel21 - Wednesday, October 16, 2019 - link

    Man, you are lazy. Reply
  • Diogene7 - Wednesday, October 16, 2019 - link

    @patel21 : How many times do you still plug an Ethernet cable to your laptop to surf on internet instead of using WIFI ? WIFI is simply more convenient...

    Similarly, wireless charging at a distance (up to ~ 1,5 foot (50cm)) would be so much more convenient than to have to plug a cable to recharge a device

    It also true for Internet of Things (IoT) devices : tjere seems to be some studies showing that consumers stop using many IoT devices that work on batteries because they have to change the batteries

    I strongly believe that wireless charging at a short distance is a requirement for the sale of IoT sensors to really take off because managing 10s or 100s or more of IoT devices with batteries is not really manageable by consumers in the long run...
    Reply
  • Molbork - Tuesday, June 16, 2020 - link

    And you just halved the power efficiency of your laptop and devices. EM transmission power is 1/r^2, checking your laptop could cost you 2-3x more than a direct connection at longer distances. Reply
  • Henk Poley - Friday, October 18, 2019 - link

    I wonder if they'll do things like heart- and breathing-rate measurement, and counting of people around you (how many hearts). Such as was demonstrated for radar based baby monitoring. Fairly low power, a 'cigarette pack' size device attached to a baby cot could work for half a year by only periodically measuring.

    Could be interesting for meetings, that your phone knows everyone has arrived, people were agitated, etc.
    Reply
  • Adonisds - Wednesday, October 16, 2019 - link

    Why is it required less than double the power to produce twice the display brightness? Reply
  • michael2k - Wednesday, October 16, 2019 - link

    Displays aren't actually perfectly transparent, and the light generating devices might absorb some of the energy instead of transmitting it.

    Increasing transparency is one way to produce more brightness with less energy.
    Reducing the amount of energy absorbed by the LEDs (and thus transformed into heat) is another way to produce more brightness with less energy.
    Changing the LEDs basic chemistry to more efficiently transform electricity into light is a third way.

    Fundamentally less waste heat, more light.
    Reply

Log in

Don't have an account? Sign up now